Stroke is currently a major contributor to disability and mortality across the globe, with ischemic stroke being the most predominant subtype. Accurate and timely diagnosis is critical for effective treatment. This study introduces a novel deep learning framework that leverages patch-level significance analysis for precise identification of ischemic strokes in Computed Tomography (CT) images. Our approach integrates a dual attention mechanism dynamic and cross attention with hybrid convolutional kernels to analyze the relative importance of brain re- gions in stroke diagnosis. The proposed architecture captures both fine-grained and contextual features to identify significant regions through attention-weighted feature embedding. The framework is evaluated on a dataset of 2023 CT of four different classes (i.e., acute: 361, chronic: 267, subacute: 382, and normal: 1013 images), employing both four and nine non-overlapping patch configurations. Experimental results demonstrate that the light gradient boosted machine classifier achieved the highest patch identification accuracy of 94.81 % and the extra tree classifier achieved an accuracy of 99.51 % for classification using 4-patch configuration analysis. The study highlights the importance of features obtained from dense layers in mitigating overfitting and improving generalization. In addition, the study reveals the potential of attention modules with interpretable factors for patch identification of cerebral infarction, suggesting the potential of artificial intelligence in aiding medical diagnosis

Dual attention mechanisms with patch-level significance embedding for ischemic stroke classification in brain CT images / Inamdar, Mahesh Anil; Gudigar, Anjan; Raghavendra, U.; Salvi, Massimo; Raj, Nithin; Pooja, J.; Hegde, Ajay; Menon, Girish R.; Rajendra Acharya, U.. - In: INFORMATICS IN MEDICINE UNLOCKED. - ISSN 2352-9148. - 57:(2025). [10.1016/j.imu.2025.101678]

Dual attention mechanisms with patch-level significance embedding for ischemic stroke classification in brain CT images

Salvi, Massimo;
2025

Abstract

Stroke is currently a major contributor to disability and mortality across the globe, with ischemic stroke being the most predominant subtype. Accurate and timely diagnosis is critical for effective treatment. This study introduces a novel deep learning framework that leverages patch-level significance analysis for precise identification of ischemic strokes in Computed Tomography (CT) images. Our approach integrates a dual attention mechanism dynamic and cross attention with hybrid convolutional kernels to analyze the relative importance of brain re- gions in stroke diagnosis. The proposed architecture captures both fine-grained and contextual features to identify significant regions through attention-weighted feature embedding. The framework is evaluated on a dataset of 2023 CT of four different classes (i.e., acute: 361, chronic: 267, subacute: 382, and normal: 1013 images), employing both four and nine non-overlapping patch configurations. Experimental results demonstrate that the light gradient boosted machine classifier achieved the highest patch identification accuracy of 94.81 % and the extra tree classifier achieved an accuracy of 99.51 % for classification using 4-patch configuration analysis. The study highlights the importance of features obtained from dense layers in mitigating overfitting and improving generalization. In addition, the study reveals the potential of attention modules with interpretable factors for patch identification of cerebral infarction, suggesting the potential of artificial intelligence in aiding medical diagnosis
File in questo prodotto:
File Dimensione Formato  
(2025) paper - brain ischemic.pdf

accesso aperto

Tipologia: 2a Post-print versione editoriale / Version of Record
Licenza: Creative commons
Dimensione 9.92 MB
Formato Adobe PDF
9.92 MB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/3002336